Claude Now Runs a Quarter of Anthropic's Own Research — and Diplomats Start Drawing Nuclear Red Lines

Two stories today sit at opposite ends of the same anxiety: how fast AI is replacing the humans who build it, and how urgently governments are trying to draw boundaries before that speed becomes irreversible.

Claude now leads more than a quarter of Anthropic's own R&D

Anthropic's new R&D Automation Index, published yesterday, reports that Claude now leads 26% of Anthropic's own AI research and development work as of August 2026 — up from under 1% in February. That's a seven-month swing from "occasionally assists" to "routinely leads," inside the very company building the model.

It's a concrete, company-reported number behind the more abstract "agents doing days of human work" claims OpenAI made earlier in the month, and it suggests the automation of AI research itself isn't a future scenario — it's already well underway at the frontier.

Three labs propose policing themselves, FINRA-style

In a direct response to the trust questions swirling since Amodei's essay, three major labs proposed a FINRA-style self-regulatory standards body for frontier AI — an industry-run organization with real enforcement teeth, modeled on how broker-dealers are overseen in finance, rather than either pure self-policing or waiting for Congress to act.

It's a pragmatic middle path, though critics will note that FINRA itself has long faced criticism for being too close to the industry it regulates.

Brookings and Fudan propose nuclear red lines, six days before Trump meets Xi

Researchers from Brookings and Fudan University jointly proposed explicit U.S.–China red lines that would bar AI from ever autonomously deciding to use nuclear weapons, alongside a dedicated military hotline specifically for AI-related incidents.

The timing is deliberate: the proposal lands six days ahead of a planned September 24 meeting between President Trump and President Xi Jinping, positioning AI-nuclear separation as one of the few areas where genuine US-China cooperation might actually be achievable, even as the two countries compete fiercely on almost everything else in AI.

OpenAI pushes into professional verticals

OpenAI launched Astra for Law, pairing GPT-6 Astra with an index of more than 230 million legal URLs; the company reports 54% correctness on a set of 200 private research questions, versus 38.7% for Astra using ordinary web search.

It's a reminder that beyond the frontier-capability and safety headlines, the quieter business of building domain-specific products on top of general models continues apace.

Anthropic opens a path for more permissive biology research

Anthropic's new Life Sciences Verification Program lets vetted research teams apply for more permissive biology safeguards across Claude and the API, with credential, security, and oversight checks built into the approval process.

It's a careful middle ground between blanket restriction and open access — letting legitimate researchers work with fewer guardrails while keeping the verification bar high enough to filter out bad actors.

The public mood keeps darkening

A new Pew survey found that more adults expect AI to cut jobs than create them in 34 of 37 countries surveyed, with the pessimistic view in the U.S. climbing seven points over the past two years.

It's consistent with the NBER productivity findings from late August — adoption keeps climbing, but public confidence that it's actually making people's working lives better keeps falling.

Between a self-regulatory body modeled on Wall Street oversight and diplomats drafting nuclear red lines six days before a presidential summit, today's news makes one thing clear: the "pace the frontier" conversation from last week didn't stay confined to corporate blog posts for long.